Automated robotic phenotyping has the potential to measure traits at a hitherto unprecedented level of detail, assisting in both crop monitoring and breeding programs for the production of new varieties. The use of 3D point cloud data allows for the measurement of geometric traits, however, occlusion results in partial point clouds. The measurement of stem length requires the computation of a skeleton through the medial axis of the point cloud, however, existing skeletonization methods are highly influenced by missing points, resulting in off-centre, non-biologically relevant skeletons. In this study, we propose a method which exploits a cylindrical shape prior during the generation of skeleton points which results in a more centred skeleton, accounting for missing points. We evaluate the k-cylinder clustering skeletonization method on real-world point clouds of strawberry petioles (stems) and demonstrate greater robustness at increased occlusion levels than the state-of-the-art.